ParisMistral says Large 4 has 1 trillion parameters, 49 billion active.
Mistral claims it significantly outperforms any open-weight model developed in the US or Europe.
It was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacentres, which also serve the preview API.
Those performance claims are Mistral’s own, and no independent evaluation has been reported.
Mistral publishes the weights by the end of October, red-teaming it with vetted partners and state authorities until then.
How each outlet framed it
- Hacker News
- emphasizes ML4 as Europe's 'AI sovereignty' challenge, cites state-of-the-art visual grounding surpassing closed-source models
Sources: Hacker News